MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement

Benchmark Model Rank Results
code-generation-on-terminal-bench-2-1MiMo-V2.6-Pro-RL#2Accuracy: 89.9
agents-on-automationbench-publicMiMo-V2.6-Pro-RL#4Success Rate: 53.1
agents-on-toolathlon-verified-108-tasksMiMo-V2.6-Pro-RL#5Pass@1: 0.769
computer-use-agents-on-osworld-verifiedMiMo-V2.6-Pro-RL#5Success Rate: 82.0
code-generation-on-terminal-bench-4-0MiMo-V2.6-Pro-RL#7Accuracy: 34.9
code-generation-on-deepsweMiMo-V2.6-Pro-RL (mini-swe-agent)#8Pass@1: 0.719